API Try Docs Blog

LFM3-VL-2B-FP8

FP8 version of LFM3-VL-2B for GPU inference.

Language feed-forward and attention weights use FP8 with 128×128 blocks and dynamic activation quantization. Embeddings, the output head, convolutions, attention gates and norms stay in BF16. The vision encoder and projector also stay in BF16. Sliding-window and RoPE settings are unchanged.

Find architecture and usage details in the original model card.

Use with SGLang

---

Support for LFM3-VL in SGLang is in review in sgl-project/sglang#46858. We will update this card when a release supports it.

python -m sglang.launch_server \
  --model-path LiquidAI/LFM3-VL-2B-FP8 \
  --reasoning-parser qwen3-thinking